Abstract / Summary
Abstract Purpose. To develop and internally validate a machine learning-based nomogram for predicting 30-day unplanned readmission after surgery for spinal metastasis, using directly measured clinical, laboratory, and operative variables from a single-institution cohort. Methods. A total of 986 patients who underwent open surgery for spinal metastasis between August 2016 and August 2026 were enrolled. Patients who died within 30 days postoperatively were excluded from the readmission analysis. Feature selection was performed using the Boruta algorithm followed by correlation analysis (excluding |r| ≥ 0.7). Six machine learning models—logistic regression (LR), random forest, gradient boosting machine, LightGBM, support vector machine, and k-nearest neighbors—were developed and compared using 10-fold cross-validation. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, Brier score, and decision curve analysis. SHAP analysis was used for model interpretation, and the final LR model was translated into a nomogram. Results. Eighty-two patients (8.32%) experienced 30-day unplanned readmission. Surgical site infection was the leading cause (26.8%). Six predictors were retained: age, number of comorbidities, preoperative albumin, operative time, intraoperative blood loss, and primary tumor pathology. The LR model achieved the highest test-set AUC of 0.905 (95% CI: 0.862–0.947), with sensitivity of 0.980, specificity of 0.765, and a Brier score of 0.062. Decision curve analysis indicated positive clinical net benefit over a clinically relevant threshold range. Conclusion. The internally validated nomogram showed strong discrimination for 30-day unplanned readmission after spinal metastasis surgery. The six predictors reflect complementary dimensions of perioperative vulnerability and may support targeted optimization aligned with enhanced-recovery principles, potentially facilitating earlier return to systemic oncological treatment. External validation is required before clinical implementation.